GPT-5.6 Luna

GPT-5.6 Luna

Quick answer: GPT-5.6 Luna is OpenAI's most cost-efficient tier in the GPT-5.6 family, released July 9, 2026. It scores 88.0% on ARC-AGI and 50.3% on Agent's Last Exam — capable frontier-level performance at just $1 input / $6 output per 1M tokens. It shares the same 1.1M-token context window as Sol and Terra.

At a Glance

Where GPT-5.6 Luna leads

• Lowest cost in the GPT-5.6 family: $1/$6 per 1M tokens — 5× cheaper than Sol
• 88.0% ARC-AGI — strong abstract reasoning for a cost-optimised tier
• 50.3% Agent's Last Exam — competitive with Terra on multi-step agentic tasks
• AA Intelligence Index 51.2, Coding Agent Index 74.6
• Full 1.1M-token context window — same as Sol and Terra
• Ideal throughput-per-dollar for high-volume production pipelines

Where it lags

• 8.5-point ARC-AGI gap vs Terra (96.5%) and 9.5 points vs Sol (97.5%)
• No max or ultra reasoning effort modes
• Lower ceiling on the highest-complexity reasoning and cybersecurity tasks
• Closed API model with no open weights

Best for: High-volume agentic pipelines, latency-sensitive applications, cost-constrained teams, and tasks where Terra or Sol capability is not required.

What GPT-5.6 Luna Is

GPT-5.6 Luna is the efficiency-first tier of OpenAI's GPT-5.6 family. It is designed for workloads where throughput, response speed, and cost per operation matter more than maximum capability. The Luna tier is a permanent capability band in the GPT-5.6 naming scheme — meaning it will improve independently as OpenAI updates the generation, without being replaced by a cheaper, weaker model.

Despite its lower price point, Luna is a capable frontier model. An 88.0% ARC-AGI score places it well above most models released before the GPT-5.6 generation, and its 50.3% Agent's Last Exam score — within 0.1 points of Terra — suggests that on multi-step agentic tasks, the practical gap between Luna and Terra may be smaller than raw benchmark numbers suggest. For many real-world agent tasks (document processing, code generation, API orchestration), Luna delivers effectively equivalent results to Terra at 60% lower cost.

Luna shares the GPT-5.6 family's full 1.1M-token context window, which distinguishes it from prior cost-tier models that typically imposed context constraints. This makes Luna a compelling option for long-context batch processing — summarising large document sets, running retrieval-augmented generation over full repositories, or analysing multi-session conversation logs — at a fraction of the per-run cost of Sol.

Specifications

Pricing

Cache writes are billed at 1.25× the uncached input rate; cache reads receive a 90% discount. Pricing per OpenAI pricing page as of July 2026.

Context Window

GPT-5.6 Luna has a ~1.1M-token context window — roughly 800–900 pages of text in a single request. This is identical to Sol and Terra, making Luna suitable for long-document and full-codebase workloads at the lowest cost per run in the GPT-5.6 family.

Public Benchmark Scores

Per-variant scores currently available for ARC-AGI and Agent's Last Exam. Shared family scores (GPQA Diamond, TerminalBench, SWE-Bench Pro, etc.) are reported for GPT-5.6 Sol — see the GPT-5.6 Sol page for the full benchmark table.

GPT-5.6 Luna vs Alternatives

Luna's strongest case vs Terra: Agent's Last Exam scores are nearly identical (50.3 vs 50.4), suggesting Luna handles multi-step agentic tasks at near-Terra level — at 60% lower cost. The larger 8.5-point ARC-AGI gap indicates Luna may underperform on novel or highly abstract single-step reasoning problems.

How GPT-5.6 Luna Performs on Real Agent Tasks

Luna's 50.3% Agent's Last Exam score is notable in context: it sits within 0.1 points of Terra (50.4%) and only 2.4 points below Sol (52.7%). For agent builders, this means Luna is a viable production tier for the large category of tasks that don't require Sol's maximum reasoning ceiling — routine code review, document summarisation, structured data extraction, and API-driven automation workflows.

Where Luna is most appropriate to evaluate first: high-volume pipelines where per-token cost directly affects unit economics; applications that run the same prompt type repeatedly (batch processing, RAG, classification); and prototyping phases where model spend should be minimised before committing to a higher tier.

Use GPT-5.6 Luna via API

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY")

response = client.chat.completions.create(
    model="gpt-5.6-luna",
    messages=[{"role": "user", "content": "Summarise the key clauses in this contract..."}],
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Compare: GPT-5.6 Sol · GPT-5.6 Terra · GPT-5.5 · GPT-5.1 Instant
• Benchmarks: ARC-AGI · Agent's Last Exam · GPQA Diamond

Frequently Asked Questions

Specs and scores sourced from OpenAI's official GPT-5.6 announcement (July 9, 2026) and Benchgen evaluations; third-party benchmark scores attributed inline. Pricing cited to the OpenAI pricing page. Last updated 2026-07-21.